Single-Channel Time-Frequency Overlapping Signal Modulation Recognition Network Based on Multi-Domain Adaptive Feature Fusion

With the application of massive wireless devices, the receiver often receives mixed signals with time-frequency overlapping. Automatic modulation classification (AMC) of such mixed signals is a critical prerequisite to effectively separating target signals. AMC algorithms typically assume that the receiver processes the single desired transmission signal. However, receiving different modulation components overlapping leads to difficulties in using the single received signal AMC algorithm during characterization. In this paper, a single-channel time-frequency overlapping signal AMC network framework based on multi-domain adaptive feature fusion (Mdaff) is designed, which utilizes the multi-dimensional adaptive feature selection (Mdafs) module to allocate the attention to extract the multi-domain features, including time-frequency, autocorrelation, and statistical domains. Then, it feeds them into the multi-scale feature fusion (Msff) recognition network based on the residual module. Multi-domain preprocessing is used to analyze the signal information comprehensively; Mdafs is used to extract important feature components from different dimensions adaptively; and the Msff module enhances the network's expressive capability. In addition, a difficulty-weighted label smoothing cross entropy (Dwlsce) function is designed to learn samples with a large degree of overlapping more adequately. We have built a software radio platform for signal acquisition, and the experimental results show that the Mdaff-based scheme has excellent AMC capability for single-channel time-frequency overlapping mixed signals.